Ear Recognition Using Elliptical Local Binary Pattern and Wavelets

See discussions, stats, and author profiles for this publication at: https://www.researchgate.net/publication/267034366

Ear Recognition using Elliptical Local Binary Patterns and Wavelets

CITATIONS

0

READS

204 3 authors:

Some of the authors of this publication are also working on these related projects:

PAPR Reduction in Multicarrier Modulation (MCM) SystemsView project

Ear BiometricView project Amir Benzaoui

Université de Bouira 24PUBLICATIONS   165CITATIONS   

SEE PROFILE

Nabil Hezil

Université 8 mai 1945 - Guelma 7PUBLICATIONS   37CITATIONS   

SEE PROFILE

Boukrouche Abdelhani Université 8 mai 1945 - Guelma 74PUBLICATIONS   307CITATIONS   

SEE PROFILE

All content following this page was uploaded by Boukrouche Abdelhani on 10 June 2019. The user has requested enhancement of the downloaded file.

Ear Recognition Using Elliptical Local Binary Pattern

and Wavelets

Amir Benzaoui*, Nabil Hezil and Abdelhani Boukrouche

Department of Electronics and Telecommunications University of 08 Mai 1945

P.O box 401, Guelma, Algeria

amirbenzaoui@gmail.com nabilhezil@gmail.com abdelhanib@yahoo.fr

Abstract—The human ear is a new technology of biometrics

which is not yet used in a real context or in commercial applications. For this purpose of biometric system, we present an improvement for ear recognition methods that use Elliptical Local Binary Pattern operator as a robust technique for characterizing the fine details of the two dimensional ear images. The improvements are focused on feature extraction and dimensionality reduction steps. The realized system is mainly appropriate for verification mode; it starts by decomposing the normalized ear image into several blocks with different resolutions. Next, the textural descriptor is applied on each decomposed block. A problem of information redundancies is appeared due to the important size of the concatenated histograms of all blocks, which has been resolved by reducing of the histogram’s dimensionalities and by selecting of the pertinent information using Haar Wavelets. Finally, the system is evaluated on the IIT Delhi Database containing two dimensional ear images and we have obtained a success rate about 94% for 500 images from 100 persons.

Keywords— biometrics, ear recognition, Local Binary Pattern, Elliptical Local Binary Pattern, texture description, haar wavelets.

I. INTRODUCTION

Biometric systems have an important role in the information and public security domains. They provide an automatic identification or verification of the identity, based on the analysis of physical or behavioral modalities of the human body. Several modalities have been used for recognizing the human identity; we can cite face, fingerprint, voice, iris, brain, computer keyboards, or signature [1, 2]. The use of the human ear in biometric applications has become a quite interesting way in last few years. It’s considered as a new class of biometrics which is not yet used in a real world or in commercial applications. The human ear is characterized by a rich structure which provides important information to differentiate between people; we can visualize 10 features and 37 sub-features from 2D ear image. The terminology of the human ear is presented in Figure 1; this terminology is made up of standard features. It includes an outer rim (helix) and ridges (antihelix) parallel to the helix, the concha (hollow part of ear), the lobe and the tragus …. [3,4,5].

The human ear has several advantages compared with others modalities; it has a rich structure, smaller object (small resolution), stable over the time, modality accepted by people, not affected by changes in age, facial expressions, position and rotation, the acquisition of the ear images can be affected without participation of the subject and can be captured from distance [3, 6].

An ear recognition system can be divided into three main steps: ear normalization, feature extraction and classification. In the normalization step, the ear image must be normalized to standard size and direction according to the long axis of outer ear contour. The long axis was defined as the line crossing through the two points which have the longest distance on the ear contour. After normalization, the long axes of different ear images were normalized to the same length and same direction [7, 8].

In feature extraction, we can recapitalize the existing methods in two principal classes, geometric and global classes. In geometric approaches, the ear analysis is given by an individual description of its parts and their relationships, like measure of distances, angles, or the triangulation between the closed edge resulting from the helix shape and lobule of -

LBP = 1 + 0 + 0 + 0 + 16 + 32 + 64 + 128 = 241 Figure 2. LBP calculation performed into 3×3 neighborhood.

- the ear [9], Crus of Helix and the Lobe [9] or the Vornoi diagram of the ear triangulation [10]. However, the methods of this approach are not effective in the unconstrained cases, i.e., situation where occlusion, lighting, or noise are uncontrolled. Recently, the scientists concentrate on global approaches, which are considered as robust approaches in the unconstrained cases compared to geometric approaches. Global approaches are based on pixel information; all pixels of the normalized ear image are treated as a single vector; the vector’s size is the total number of the pixels. Principal Component Analysis (PCA) [11], Color Spaces fusion [12], 2D Gabor Filter [13] … are the most popular methods used to change the space of representation, to change the data representation, to reduce the dimensionalities, or to select only the useful information.

This paper presents an improvement for automated ear recognition systems and comparison with last state of the art approaches. There are four steps of the proposed algorithm. First, preprocessing is applied to improve the contrast and the quality of the images. Second, the ear image is decomposed into several blocks with different resolutions. Next, the textural descriptor ELBP is applied on all blocks of the decomposed image. Finally, Haar Wavelets in one dimensional space are used to reduce the dimensionalities of the concatenated histograms from each block.

This paper is organized as follows: in the next Section, we describe the classical LBP and ELBP. In the Section 3, the enrolment phase and ear characteristic extraction is presented. In the Section 4, we present the verification phase and performances evaluation applied on IIT Delhi database. For this purpose, Chi-Square distance is required to measure similarities between two ear templates. Finally, a conclusion related to this work is given in Section 5.

II. LOCAL BINARY PATTERN (LBP)

The original LBP operator introduced by Ojala et al. [14], which has been used for texture discrimination, has shown a powerful and effective results against the variations in rotation and illumination. The operator labels the pixels of an image by thresholding the 3×3 neighborhood of each pixel with the

central value and considering the result as a binary code. Next, the histogram of the labels can be used as a texture descriptor

(see Figure 2); for a given a pixel gc(xc,yc) from gray image,

its texture LBP is calculated by comparing gc with its

neighbors pixels P on a circle of radius R (see Figure 3 for

more details on circular neighborhood). The value of LBP(gc)

is obtained as:

LBPP,R_{(xc, yc) = ∑Pi=1 S(giP ,R - gc) 2}i-1 ₍₁₎

S(x) is defined as: S x) = _{0 otherwise;}1 if x ≥ 0; (2) However, the major disadvantage of the original LBP operator resides in the size of the descriptor, a mask of 3×3 pixels cannot capture the structures of large scales which can considered as a dominant structures in the image. Recently, the size of the operator has been extended by using a mask with different large sizes. Figures 3.a, 3.b, 3.c show three examples of the extended LBP.

A new model of the extended operators of LBP called: Elliptical Local Binary Patterns (ELBP) [15] which presents significant improvements for encoding the micro features (the fine details) of the image compared with LBP. In ELBP, at

each pixel gc(xc,yc), we consider its surrounding pixels that lie

on an ellipse (see Figure 4) with (xc,yc) is the center. ELBP of

(xc, yc) with P neighboring pixels at (R1, R2) distances is

computed as:

ELBPP,R1,R2_{(xc, yc) = ∑Pi=1 S(giP ,R1,R2 - gc) 2}i-1 ₍₃₎

S(x) function is defined as (2).

(a) (b) (d)

Figure 3. Neighborhood set for different (P,R). (a) The basic LBP operator (P,R) = (8,1) (b) LBP with circular neighborhood (8,2). (c) LBP with circular

neighborhood (8,3). 6 5 2 7 6 1 13 28 11 1 0 0 1 0 1 1 1 1 2 4 128 8 64 32 16 1 0 0 128 241 0 64 32 16 Threshold _Multiplay

ELBP (8,2,3) ELBP (8,3,2) ELBP (8,3,4) Figure 4. ELBP samples with different extention [15].

Figure 5. Results of LBP and ELBP application with differentd masks.

In details, the coordinates of the ith neighboring pixel of (xc,yc)

are calculated using the formulas:

angle-step = 2 * π / P (4) xi = xc + R1 * cos (( i – 1) * angle-step) (5)

yi = yc + R2 * sin (( i – 1) * angle-step) (6)

Figure 5 shows the results of LBP and ELBP applications using different masks. However, the extend versions of LBP operators and the Elliptical LBPs present a good results by capturing the local patterns and the micro features of the human ear but they are not performed for capturing the global characteristics which can considered as dominant structures in the image.

III. PROPOSED APPROACH

The proposed algorithm used to extract information for ear recognition is described in the following recapitulation; next we present each step in details. We consider that we have a gallery θ of biometric samples with P persons, S biometric

sample (image) per person (example θ=500, P=100, S=5, with

IIT Delhi database). The process of features extraction is composed of five principal stages for each person:

a. Preprocessing of all normalized ear images.

b. Decomposition of each ear image into several blocks

with different resolutions.

c. Application of the textural descriptor ELBP for each

decomposed block.

d. Concatenation of the resulting histograms from each

block in one global histogram.

e. Dimensionalities reduction of each global histogram

using Haar Wavelets. A. Preprocessing

The objective of the preprocessing is the modification of the source’s image representation to facilitate the task of the following steps and to improve the rate of recognition. First, the ear image is converted into grayscale image. Next, every grayscale image is filtered by median filter to suppress noise. Lastly, the noise suppression image is then adjusted to improve the contrast of the image [16].

B. Image decomposition

Most LBP operators characterize the texture distribution of each pixel with its neighborhood only. But, the differences between two ear images cannot demonstrate by the texture distribution of each pixel with its neighborhood only, but also by the relative connection with other pixels in the image. With this intention, we have decomposed the original image into several sub-images (see Figure 6) to characterize better the details and the relationships between all pixels of the image. In the next stage, the extracted histograms from each block will be concatenated in one global histogram. With this technique, we can obtain the fine details and the relative connections between all pixels of the image.

Figure 6. Image decomposition in differents blocks. ELBP (8,3,4) ELBP (8,3,2) LBP (8,2) Original Image LBP (8,1) Level 1 Level2 Level3

C. Application of the textural descriptor ELBP and histogram’s concatenation

The textural descriptor ELBP, presented in the previous section, is applied on all blocks of the decomposed image with the different resolutions like presented in Figure 6. Next, the extracted histograms from each block will be concatenated in one global histogram (vector) representing an ear template. A problem of information redundancies is appeared due to the important size of each template.

D. Dimensionalities reduction using Haar Wavelets

To resolve the problem of information redundancies, we have used the Discrete Wavelet Transform (DWT) as a technique of data compression, to reduce the dimensionalities and to select only the useful information needed to model each person. Wavelet analysis is the breaking up of a signal into a set of scaled and translated versions of an original (or mother) wavelet. Taking the wavelet transform of a signal decomposes the original signal into wavelets coefficients at different scales and positions. These coefficients represent the signal in the wavelet domain and all data operations can be performed using just the corresponding wavelet coefficients.

Wavelets work by decomposing the concatenated histograms (global histogram) into different resolutions or frequency bands, and choosing the Haar Wavelet and computing the 1D Discrete Wavelet Transform (DWT). The extraction of pertinent information is based on the concept of selecting a small number of approximated coefficients (at a suitably chosen level) and some of the detail coefficients can accurately represent regular signal components. Choosing a decomposition level for the DWT usually depends on the type of signal being analyzed or some other suitable criterion such as entropy. For the processing of ear templates decomposition in our application up to scale 3 is adequate.

III. VERIFICATION PHASE AND PERFORMANCES

EVALUATIONS

For a given image ξ considered as a test example in the verification mode. In this case, ξ is declared as present in the database or unknown. We start by building its template representation, related to individual’s ear, using the extraction

method presented in this work (stages: a-e). This template is

compared to the stored templates of the same person in the Gallery using the Chi-Square distance. This distance between two templates is calculated as:

Distchi-square (X, Y) = ∑Mi=0(xi −yi )2_{xi +yi} (8)

Where xi and yi are the two bit-wise codes to be compared.

To evaluate the performances of the realized system, we have carried out several tests on IIT Delhi database. We have taken 500 normalized ear images for evaluation. The IIT Delhi ear image database consists of the ear image database collected from the students and staff at IIT Delhi, New Delhi, India. This database has been acquired in IIT Delhi campus

during Oct 2006 - Jun 2007 using a simple imaging setup. All the images are acquired from a distance using simple imaging setup and the imaging is performed in the indoor environment. The currently available database is acquired from the 121 different subjects and each subject has at least three ear images. All the subjects in the database are in the age group 14-58 years. The resolution of these images is 272×204 pixels and all these images are available in jpeg format (Figure 7). In addition to the original images, this database also provides the automatically normalized and cropped ear images of size 50×180 pixels (Figure 8) [13, 17].

In this evaluation, 3 images for each person are used as Gallery and the others2 imagesare used as the probe-sets. The Gallery and the probe-sets separately consist of 300 and 200 images. The proposed approach has been evaluated with all possible probabilities and the mean rate of all possible selections has taken in consideration.

As it is well known, to validate any biometric verification, it is essential to calculate the False Acceptance Rate (FAR) and the False Rejection Rate (FRR) to find the accuracy of the biometric system which is calculated as:

Accuracy = 100 − (𝐹𝐴𝑅 +𝐹𝑅𝑅)₂ (7) The last rate is known under the name of: Equal Error Rate (EER). This rate is calculated from the first two rates, it constitutes a current point of performance’s measurement (The center of gravity between intra-class and inter-class). This rate corresponds to the point where FRR = FAR, i.e., the compromise between the False Acceptance Rate and the False Rejection Rate.

 If Dist <= EER: this indicates that two 1D ear templates

are very close and can be considered that they correspond to the same person authentic.

 Else, if Dist > EER: two 1D codes are considered in this

case different, which can be used to distinguish between two persons.

Figure 7. Some images from IIT Delhi database [13,17].

TABLE I. COMPARISON WITH OTHER STATE OF ART APPROCHES USING THE SAME DISTRIBUTION

Approach Accuracy %

PCA without decomposition 80,39

Classical LBP 89 Classical LBP + PCA 92,75 Classical LBP + DWT 92,71 LBP (8,2) 89,8 LBP (8,3) 86,14 LBP (8,2) + PCA 90,43 ELBP(8,2,3) 91,4 ELBP(8,3,2) 91,63 ELBP(8,3,4) 91,39 ELBP(8,3,2) + PCA 93,44 ELBP(8,3,2) + DWT 94,51

We have compared also the proposed approach with other state of art approaches using the similar protocols under more challenges and scenarios. The ear images are taken with IIT Delhi database which presents a very good variation, in lighting and occlusion, to measure the performances of the approach in the difficult situations. The results of the comparison are presented in Table1.

Table 1 shows clearly that ELBP presents very good results with IIT Delhi database which indicates that this descriptor has an important effectiveness against to the variations in different factors like as occlusion, lighting, rotation, and noise.

The comparison results from Table 1 show that our ELBP (8,3,2) + DWT method outperforms other state of the art systems using the same distribution. These results also reveal that ELBP (8,3,2), in normalized ear modality, is more robust than LBP in the extraction of the fine details (micro features). Another conclusion we can make from the Table 1 is that ELBP DWT is much better than ELBP PCA; the association of the DWT as a robust technique in the dimensionalities reduction is very interesting to improve the performances of the realized approach.

V. CONCLUSION

In this work, we have successfully developed a simple approach for ear texture discrimination; this approach is primary based on Elliptical Local Binary Patterns. Each normalized ear image is decomposed on multi blocks with different resolution, and the textural descriptor ELBP will be applied on each decomposed block. Next, the extracted histograms from each block will be concatenated in one global histogram. Finally, DWT Wavelets is applied on each global

histogram to reduce the dimensionalities and to extract the useful information. The experimental results applied on IIT Delhi database have showed that the feature extraction approach based on ELBP and DWT has given a very significant improvement at the recognition rate.

REFERENCES

[1] A.K. Jain, A.A. Ross and K. Nandakumar., “Introduction to Biometrics,” 1st _{edn. Springer US, 2011.}

[2] A.K. Jain, P. Flynn and A.A. Ross., “Handbook of Biometrics”, 1st _edn.

Springer US, 2008.

[3] Arbab-Zavar B. and Nixon M.S., “On guided model-based analysis for ear biometrics,” Computer Vision and Image Understanding (Elsevier), vol.115, pp.487-502, 2011.

[4] Kumar A. and Tak-Shing T.C., “Robust ear identification using sparse representation of local texture descriptors,” Pattern Recognition (Elsevier), vol.46, pp.73-85, 2013.

[5] Dinkar A.D. and Sambyal S.S., “Person identification in Ethnic Indian Goans using ear biometrics and neural networks,” Forensic Science International (Elsevier), vol.223, pp.373.e1-373.e13, 2012.

[6] Yuan L. and Mu Z.C., “Ear recognition based on local information fusion,” Pattern Recognition Letters (Elsevier), vol.33, pp.182-190, 2012.

[7] Yuan L. and Mu Z.C., “Ear recognition based 2D Images,” Proceedings of the 1st _{IEEE International Conference on Biometrics: Theory,}

Applications, and Systems (BTAS), Washington, USA, September 27-29, 2007.

[8] Xie Z. and Mu Z.C., “Ear recognition Using LLE and IDLLE Algorithm,” Proceedings of the 19th _{IEEE International Conference on}

Pattern Recognition (ICPR), Florida, USA, December 08-11, 2008. [9] Burge M. and Burger W., “Ear Biometrics in Computer Vision,”

Proceeding of the 15th _{IEEE International Conference on Pattern}

Recognition (ICPR), Barcelona, Spain, pp.822-826, September 3-8, 2000.

[10] Burge M. and Burger W., “Ear biometrics in machine vision,” Proceedings of the 21st _{Workshop of the Australian Association for}

Pattern Recognition, 1997.

[11] Arbab-Zavar B., Nixon MS. And Hurley D.J., “On model-based analysis of ear biometrics,” Proceedings of the 1st _{IEEE International Conference}

on Biometrics: Theory, Applications, and Systems (BTAS), Washington, USA, September 27-29, 2007.

[12] Nanni L. and Lumini A., “Fusion of color spaces for ear authentication,” Pattern Recognition (Elsevier), vol.42, pp.1906-1913, 2009.

[13] Kumar A. and Wu C., “Automated human identification using ear imaging,” Pattern Recognition (Elsevier), vol.45, pp.956-968, 2012. [14] Ojala T., Pietikainen M. and Harwood D., “A Comparative Study of

Texture Measures with Classification based on Features Distribution,” Pattern Recognition (Elsevier), vol.29, no.06, pp.51-59, 1996.

[15] H.T. Nguyen and A. Caplier., “Elliptical Local Binary Pattern for Face Recognition,”Springer: Proceeding ofInternational Workshops on Computer Vision (ACCV’12), Daejeon, Korea, pp.896, November 5-6, 2012.

[16] Benzaoui A., Bourouba H., and Boukrouche A., “System for Automatic Faces Detection,” Proceeding of the 3rd _{IEEE International Conference}

on Image Processing, Theory, Tools and Applications (IPTA’3), Istanbul, Turkey, pp.354-358, Octobre 15-18, 2012.

[17] IIT Delhi Ear Database, Version 1:

http://www4.comp.polyu.edu.hk/~csajaykr/IITD/Database_Ear.htm

View publication stats View publication stats